Cohere's North Mini Code: An Open 30B MoE Model for Agentic Coding

Cohere has released North Mini Code, the company's first open model for developers, built on a mixture-of-experts architecture. Of its 30 billion parameters, the model activates only 3 billion on each call. It fits on a single H100 GPU and supports a context window of 256 thousand tokens.

Agentic coding as the main use case

North Mini Code was built specifically for agentic coding, where the AI does not wait for a prompt but plans, writes, tests and fixes code on its own in a loop. This is a different class of task from ordinary autocomplete: the model receives a high-level assignment and solves it with minimal developer intervention.

The 256-thousand-token context window is an architectural choice made with these scenarios in mind. It can hold entire repositories in memory: not one file, but dozens of interrelated modules, tests, configs and documentation at the same time. For end-to-end refactoring or tracing a bug's root cause through several layers of abstraction, this is essential.

Typical tasks the model is optimized for:

• Autonomous refactoring that accounts for the whole codebase
• Multi-step debugging without constant prompting
• Analysis and documentation of legacy projects
• CI/CD integration as an AI reviewer
• Generating and running test cases in an automated loop

How MoE works in practice

Mixture-of-experts is an approach in which the model is split into specialized expert blocks, and only a subset of them is activated for each token. Of North Mini Code's 30 billion parameters, about 3 billion are in use at any moment, roughly a tenth.

In practice, this means the compute load is closer to that of a 3-billion-parameter model, while response quality is closer to that of a 30-billion-parameter one. The model runs on a single NVIDIA H100 GPU, which sharply lowers the barrier to deployment. A rented H100 in the cloud costs $2–4 per hour, versus tens of thousands for a cluster running an equivalent dense model.

MoE architectures are becoming the standard for efficient scaling: DeepSeek-V3 and Mixtral took the same path and showed that a model can compete with larger dense models at lower inference cost.

Open weights as a competitive argument

Cohere is publishing North Mini Code with open weights. That means you can download the model, fine-tune it on your own data and deploy it on any infrastructure, with no dependence on the Cohere API. For enterprise customers, this covers several key needs: data stays inside the company perimeter, the model can be adapted to internal code standards and proprietary frameworks, and latency with local deployment is minimal.

This demand for control, customizability and independence from SaaS is becoming Cohere's main competitive argument against closed solutions.

What this means

The market for AI developer models is crowded: DeepSeek Coder, Qwen2.5-Coder, Code Llama, Mistral and others are already in it. North Mini Code targets a specific niche, agentic tasks on a company's own infrastructure, and bets on three things at once: MoE efficiency, long context and open weights.

Whether that combination pays off will be shown by independent benchmarks and the first production deployments.